Predicted Gene Expression Tested in Partridge Chicken Genomic Selection Models

Integrating computationally predicted gene expression levels into standard genomic prediction models yielded numerical gains in predicting body weight traits in Tiannong partridge chickens, but these improvements did not reach statistical significance after multiple-testing correction, according to an evaluation published in Genetics Selection Evolution. The investigation, led by researchers including Huimin Kang, Weicong Wu, Hua Li, and colleagues, evaluated whether imputing genetically regulated expression could enhance genomic estimated breeding value predictions in agricultural scenarios where direct RNA sequencing across validation animals is absent or cost-prohibitive. Across realistic breeding configurations, incorporating predicted expression estimated breeding values produced modest numerical increases in prediction accuracy over conventional genomic best linear unbiased prediction, yet those differences remained statistically indistinguishable from baseline genomic models.

Genomic prediction fundamentally relies on dense single nucleotide polymorphism markers distributed across the genome to calculate genomic estimated breeding values, bypassing the requirement of traditional marker-assisted selection to locate specific quantitative trait loci. While widely implemented across cattle, swine, and poultry breeding to accelerate genetic gain, standard statistical formulations like genomic best linear unbiased prediction predominantly rely on linear marker relationships and omit intermediate functional biological layers. Because transcriptomic variation sits functionally closer to complex traits within biological pathways than raw sequence variation, incorporating expression data represents an attractive strategy to improve selection accuracy. However, generating whole-transcriptome sequencing across routine commercial flocks remains financially prohibitive, creating a persistent bottleneck that restricts multi-omics evaluations to small cohorts.

To navigate sequencing cost limitations, the research team leveraged the principle that gene expression acts as a heritable quantitative trait governed in part by local genetic architecture. Using genotype-tissue expression methodologies expanded through efforts such as the FarmGTEx initiative, expression estimated breeding values can be derived exclusively from single nucleotide polymorphism profiles once training models are established. The investigators examined three hundred and fifty healthy female Tiannong partridge chickens, a commercially utilized three-way cross derived from indigenous Chinese Qingyuan partridge stock known for slow growth and specific meat characteristics. All birds were hatched in a single cohort, maintained under uniform commercial housing and nutritional regimens through day one hundred and twenty-five, and slaughtered on day one hundred and twenty-six to record live weight, dressed weight, and eviscerated weight under standardized criteria.

Genomic DNA isolated from blood samples was subjected to whole-genome resequencing at an average sequencing depth of fourfold per bird, aligned to the GRCg6a chicken reference assembly, and filtered through quality control protocols to yield two million one hundred and thirty-eight thousand one hundred and fifty-seven high-quality autosomal single nucleotide polymorphisms across all three hundred and fifty birds. To prevent information loss, the researchers deliberately retained all quality-filtered autosomal markers without linkage disequilibrium pruning, as preliminary cross-validations revealed that pruning consistently reduced prediction accuracy. Simultaneously, total RNA extracted from snap-frozen right pectoral muscle collected at slaughter was sequenced using paired-end one hundred and fifty base-pair sequencing on an Illumina platform, establishing normalized transcript levels across twelve thousand six hundred and ninety-one expressed genes.

To rigorously assess multi-omics integration, the authors formulated three distinct experimental scenarios evaluated through ten independent repetitions of five-fold cross-validation. In Scenario I, observed pectoral muscle transcriptomic profiles were available for both the reference and validation sets. In Scenario II, observed transcriptomic data were limited strictly to the reference population, requiring models trained on reference genotypes and transcriptomes to predict validation gene expression breeding values. In Scenario III, simulating an external reference resource, neither the reference nor the validation group possessed observed RNA data; instead, the external ChickenGTEx muscle panel of five hundred and seventeen birds was utilized to train expression prediction models using two hundred and eighty-nine thousand seven hundred and thirty common single nucleotide polymorphisms.

In Scenario I, where observed tissue expression values were directly incorporated into genomic-transcriptomic best linear unbiased prediction, the model demonstrated statistically significant superior performance over baseline genomic prediction. Holm-corrected paired t-tests yielded adjusted p-values below zero point zero zero zero one for live weight and eviscerated weight, and zero point zero zero five for dressed weight. However, the authors noted that this observed-expression model captures both heritable genetic variation and non-genetic environmental influences, thereby predicting overall phenotypic performance rather than purely transmissible breeding values. Conversely, a two-step transcriptomic-genomic best linear unbiased prediction designed to isolate the additive genetic component yielded prediction correlations of zero point one six seven for live weight, zero point two one four for dressed weight, and zero point two one seven for eviscerated weight, which were not statistically different from standard genomic BLUP baselines of zero point one seven eight, zero point two one seven, and zero point one nine three.

When examining cis-heritability distributions for pectoral gene expression, the researchers observed a pronounced right-skewed pattern in both Scenario II and Scenario III, with the vast majority of genes displaying low cis-heritability estimates alongside a sparse subset of highly heritable transcripts. Expression estimated breeding value predictions were modeled using either best linear unbiased prediction or penalized elastic net regression across single nucleotide polymorphisms located within one megabase of transcription start sites. The elastic net algorithm generally surpassed linear unbiased prediction across Scenarios II and III, showing a positive correlation between prediction accuracy and cis-heritability. Nevertheless, in Scenario III, imposing an overly stringent cis-heritability threshold above zero point seven reduced the number of retained genes to just three, introducing substantial estimation uncertainty.

In Scenario II, where expression values were predicted internally for the validation set, the elastic net model filtering for genes with cis-heritability exceeding zero point seven produced the highest numerical genomic prediction accuracies, reaching zero point one eight nine for live weight, zero point two three one for dressed weight, and zero point two zero three for eviscerated weight. These values represented relative gains of six point two percent, six point five percent, and five point two percent over baseline genomic prediction, respectively. Similarly, under the cross-population framework of Scenario III, the elastic net model retaining genes with heritability exceeding zero point three attained relative numerical gains over genomic BLUP of thirteen point five percent for live weight, seventeen point five percent for dressed weight, and nine point three percent for eviscerated weight. However, every Holm-adjusted pairwise comparison against baseline genomic BLUP yielded adjusted p-values of one point zero zero, indicating an absence of formal statistical significance.

The study also examined genomic-feature best linear unbiased prediction, a feature-selection strategy that partitions markers into background variants and functional variants identified through expression quantitative trait locus mapping. In Scenario II, trans-eQTL feature modeling produced only an unconfirmed zero point seven percent numerical increase for dressed weight, while cis-eQTL modeling showed no advantage. In Scenario III, transferring regulatory features derived from the external ChickenGTEx dataset caused genomic-feature models to perform numerically worse than standard genomic BLUP across all three measured traits. This performance decline highlights the pronounced biological difficulty of transferring fine-scale regulatory variants across distinct genetic populations characterized by differing linkage disequilibrium structures, divergence in allele frequencies, and genotype-by-environment interactions.

Reflecting on the findings, the authors emphasized that the relatively small reference cohort of three hundred and fifty birds restricted the statistical power required to detect modest differences between complex prediction frameworks. While predicted gene expression estimated breeding values offer a biologically grounded intermediate layer that circumvents individual-level RNA sequencing in target birds, their practical advantage over standard single nucleotide polymorphism evaluations remains numerically suggestive rather than statistically verified. The researchers concluded that confirming whether intermediate transcriptomic imputation can provide reproducible, economically sustainable advantages in commercial animal improvement will require extensive testing across larger, genetically diverse populations alongside refined cross-population statistical methodologies.

Subject of Research: Genomic selection and predicted gene expression integration in poultry breeding

Article Title: Integrating predicted gene expression into genomic selection: an evaluation in partridge chickens under realistic breeding scenarios

Article References: Kang, H., Wu, W., Ma, Z., Yang, T., Xiang, H., Ye, F., Yu, H., & Li, H. (2026). Integrating predicted gene expression into genomic selection: an evaluation in partridge chickens under realistic breeding scenarios. Genetics Selection Evolution, 58(1), Article 83. https://doi.org/10.1186/s12711-026-01099-w

Image Credits: AI Generated

DOI: 10.1186/s12711-026-01099-w

Keywords: genomic prediction, partridge chickens, gene expression, GBLUP, GTBLUP, poultry breeding, transcriptomics, multi-omics, Integrating, predicted, gene, expression